Hybrid 1D convolutional neural network and fuzzy clustering for damage detection in composite plates using laser ultrasonic guided waves

This paper introduces a hybrid learning framework for structural health monitoring (SHM) of composite laminates using laser ultrasonic guided waves (LUGWs) with limited labelled data. A lightweight one-dimensional convolutional neural network (1D-CNN) is trained on a small, spatially labelled subset to extract low-dimensional latent features from raw waveforms. These features are clustered using five classical algorithms (K-Means, GMM, BGMM, DBSCAN, and HDBSCAN) and a novel method, namely Probabilistic CNN–Fuzzy Clustering (PCFC). PCFC integrates CNN-derived class probabilities with fuzzy membership values from the Fuzzy C-Means algorithm, thereby enabling soft, uncertainty-aware clustering in the latent space. The quality of the clustering is evaluated through spatial cluster map reconstructions, with PCFC consistently producing coherent and physically meaningful clusters that are aligned with known delamination regions. In the latent feature space, Principal Component Analysis reveals a consistent separation between damaged and undamaged waveforms along the first principal component, thus offering an interpretable indicator of damage severity. The proposed approach has the potential to reduces annotation requirements by an order of magnitude and mitigate the limitations of fully supervised learning. It offers a scalable and interpretable solution for reliable damage detection in composite structures, particularly in scenarios where comprehensive labelling is impractical.